摘要
新一轮科技发展与全球贸易发展推动智慧海关建设,而通用大模型在海关领域存在专业理解不足、准确性有限等问题,难以满足精准监管与快速通关需求。本研究旨在通过生成式大模型技术赋能智慧海关,构建协同高效的现代化海关治理体系。研究以生成式大模型为核心,整合海关商品细则、关税政策等多类文档构建向量库,设计“数据增强模型-模型赋能海关”的闭环架构,研发单据数据比对、精准知识生成两大核心算法,形成多终端、多模态的知识生成系统。其核心技术价值体现在处理多模态数据破解监管信息差、模拟人类推理解决复杂归类决策、优化官员多轮对话交互实现自然语言服务三方面。通过对国产生成式大模型进行领域微调、知识增强及并行预测优化,研究打造了覆盖商品归类、易混淆商品辨析、海关知识智能问答、大数据风险研判等场景的应用系统。该系统已在天津海关相关场景试点应用,有效提升了监管效率与准确性,为智慧海关信息化建设提供了关键技术支持。未来,随着技术深度融合与国际合作推进,将进一步支撑贸易便利化与监管现代化目标实现。
The new round of technological advancements and global trade development has driven the construction of smart customs systems.However,general-purpose large models face challenges such as insufficient domain expertise and limited accuracy in the customs field,making it difficult to meet the demands of precise supervision and rapid clearance.This study aims to empower smart customs with generative large model technology,establishing a collaborative and efficient modern customs governance system.The research focuses on generative large models,integrating various documents such as customs commodity regulations and tariff policies to build a vector database.It designs a closed-loop architecture of“data-augmented models empowering customs”,develops two core algorithms for document data comparison and precise knowledge generation,and forms a multi-terminal,multimodal knowledge generation system ecosystem.The core technical value lies in three aspects:processing multimodal data to bridge regulatory information gaps,simulating human reasoning to resolve complex classification decisions,and optimizing customs officers′multi-turn dialogue interactions to achieve natural language services.Through domain fine-tuning,knowledge enhancement,and parallel prediction optimization of domestic generative large models,the study has developed an application system covering scenarios such as commodity classification,differentiation of easily confused goods,intelligent customs knowledge Q&A,and big data risk assessment.This system has been piloted in relevant scenarios at Tianjin Customs,effectively improving supervision efficiency and accuracy while providing critical technical support for smart customs informatization.In the future,with deeper technological integration and international collaboration,it will further support the goals of trade facilitation and modernized supervision.
作者
张程
李彤彤
吕哲楠
张弛
张宇含
张成喆
陈婧琦
姜懿轩
牟华
ZHANG Cheng;LI Tongtong;LV Zhenan;ZHANG Chi;ZHANG Yuhan;ZHANG Chengzhe;CHEN Jingqi;JIANG Yixuan;MU Hua(Tianjin Customs,Tianjin 300012,China;China Electronic Port Data Center Tianjin Branch,Tianjin 300012,China)
出处
《口岸非传统安全学刊》
2025年第6期65-71,共7页
JOURNAL OF NON-TRADITIONAL BORDER SECURITY SCIENCE AND TECHNOLOGY
基金
海关总署科技项目(2024HK302)。
关键词
生成式大模型
智慧海关
知识增强
海关监管
多轮对话交互
generative large models
smart customs
knowledge augmentation
customs supervision
multi-turn dialogue interaction
